bytedance/deer-flow: Deep Research and Content Skills for Claude Code

What bytedance/deer-flow ships

Flowy indexes 32 components for bytedance/deer-flow: 29 skills and 3 agents, from a repository with 81,051 GitHub stars. It is filed under research: most of it runs from a question to a document, chart, or media file, plus a separate set of tools for building and grading skills. Its README notes DeerFlow reached number one on GitHub Trending on February 28th, 2026, after shipping version 2. The install command and a link to the source repository are on the listing page linked at the top of this guide.

Which skills handle the research itself

Seven skills form the research core. deep-research replaces an ad hoc web search: use it instead of WebSearch for any question needing research, with a multi-angle method instead of one search and a summary. academic-paper-review reviews and critiques academic papers and preprints. systematic-literature-review runs a survey across multiple papers on arXiv, producing an annotated bibliography or comparison in APA, IEEE, or BibTeX format. github-deep-research runs the same kind of multi-round investigation against a GitHub repository, reconstructing a timeline or competitive analysis. consulting-analysis writes market, brand, financial, industry, or investment due diligence reports. data-analysis works on Excel and CSV files: multi-sheet workbooks, pivot tables, and SQL-style queries. claude-to-deerflow calls the DeerFlow platform's own HTTP API instead: start a conversation thread, check status, or list its models, skills, and agents, useful only if you also run DeerFlow as a hosted service.

Which skills turn findings into something you can hand someone

Eight skills cover output. chart-visualization picks from 26 chart types and renders the image with a JavaScript script. image-generation and video-generation both work from a structured prompt and an optional reference image. music-generation composes a song from a style or mood prompt through the MiniMax music API. podcast-generation turns written content into two-host conversational audio. ppt-generation builds a slide deck by generating an image per slide and assembling a PowerPoint file. newsletter-generation researches a topic and formats the result as an email digest or industry briefing. code-documentation closes the group: README files, API references, comments, architecture docs, and changelogs, documenting code rather than research.

What frontend-design and web-design-guidelines add

Two skills sit outside both groups above. frontend-design builds interfaces, components, pages, and artifacts rather than research output. web-design-guidelines reviews UI code someone already wrote against Web Interface Guidelines and accessibility practice. Neither depends on the research skills.

The skill-building tools, plus two outliers

A third cluster is about skills themselves. skill-creator creates, edits, and benchmarks skills, including running evals with variance analysis. skill-reviewer audits a finished skill package for readiness: its triggers, safety boundaries, resources, and evidence. find-skills locates an existing skill instead of writing a new one, and surprise-me discovers whatever skills are enabled and combines them for an unplanned demo. Two more fit neither theme: bootstrap runs an onboarding conversation to generate a personalized identity file for an AI assistant, and vercel-deploy-claimable deploys an app or website to Vercel, needing no authentication, by its own description.

The plugin's 3 agents back this cluster. comparator compares two outputs without being told which skill produced either one. analyzer studies that blind comparison to explain why the winner won and suggests improvements. grader checks a transcript and its outputs against a stated set of expectations. Together they read as the evaluation harness that skill-creator's benchmarking and skill-reviewer's audits run on.

Why six skill names read like test fixtures, not features

Six of the 29 skills, blocked, needs-revision, partial-package, prompt-injection, publish-candidate, and zh-output, sit together in the indexed list right after skill-reviewer, and their own descriptions explain why. partial-package calls itself a fixture for truncated package review. prompt-injection says it demonstrates unsafe review data for calibration tests. blocked has a placeholder description standing in for a missing one. needs-revision, publish-candidate, and zh-output skip the word fixture, but match the same pattern of review verdicts and edge cases rather than a task someone would ask for, so they read as the same calibration set for skill-reviewer. If you install this plugin expecting 29 tools to reach for directly, expect six of them to be test data instead.

Who this plugin fits

This fits someone who wants one plugin for a research question, web search, academic paper, GitHub repository, or spreadsheet, then wants the answer packaged as a chart, slide deck, newsletter, or audio file, instead of switching plugins per format. It also fits someone building their own Claude Code skills who wants a comparator, an analyzer, and a grader on hand instead of judging outputs by eye. It fits less well if you only want one narrow thing: a plugin built around a single chart type or document format does less, and it will not hand you six skills you need to filter out first.